[SPARK-17219][ML] enhanced NaN value handling in Bucketizer

## What changes were proposed in this pull request?

This PR is an enhancement of PR with commit ID:57dc326bd00cf0a49da971e9c573c48ae28acaa2.
NaN is a special type of value which is commonly seen as invalid. But We find that there are certain cases where NaN are also valuable, thus need special handling. We provided user when dealing NaN values with 3 options, to either reserve an extra bucket for NaN values, or remove the NaN values, or report an error, by setting handleNaN "keep", "skip", or "error"(default) respectively.

'''Before:
val bucketizer: Bucketizer = new Bucketizer()
          .setInputCol("feature")
          .setOutputCol("result")
          .setSplits(splits)
'''After:
val bucketizer: Bucketizer = new Bucketizer()
          .setInputCol("feature")
          .setOutputCol("result")
          .setSplits(splits)
          .setHandleNaN("keep")

## How was this patch tested?
Tests added in QuantileDiscretizerSuite, BucketizerSuite and DataFrameStatSuite

Signed-off-by: VinceShieh <vincent.xieintel.com>

Author: VinceShieh <vincent.xie@intel.com>
Author: Vincent Xie <vincent.xie@intel.com>
Author: Joseph K. Bradley <joseph@databricks.com>

Closes #15428 from VinceShieh/spark-17219_followup.

docs/ml-features.md

@@ -1103,11 +1103,16 @@ for more details on the API.
 
 `QuantileDiscretizer` takes a column with continuous features and outputs a column with binned
 categorical features. The number of bins is set by the `numBuckets` parameter. It is possible
-that the number of buckets used will be less than this value, for example, if there are too few
-distinct values of the input to create enough distinct quantiles. Note also that NaN values are
-handled specially and placed into their own bucket. For example, if 4 buckets are used, then
-non-NaN data will be put into buckets[0-3], but NaNs will be counted in a special bucket[4].
-The bin ranges are chosen using an approximate algorithm (see the documentation for
+that the number of buckets used will be smaller than this value, for example, if there are too few
+distinct values of the input to create enough distinct quantiles.
+
+NaN values: Note also that QuantileDiscretizer
+will raise an error when it finds NaN values in the dataset, but the user can also choose to either
+keep or remove NaN values within the dataset by setting `handleInvalid`. If the user chooses to keep
+NaN values, they will be handled specially and placed into their own bucket, for example, if 4 buckets
+are used, then non-NaN data will be put into buckets[0-3], but NaNs will be counted in a special bucket[4].
+
+Algorithm: The bin ranges are chosen using an approximate algorithm (see the documentation for
 [approxQuantile](api/scala/index.html#org.apache.spark.sql.DataFrameStatFunctions) for a
 detailed description). The precision of the approximation can be controlled with the
 `relativeError` parameter. When set to zero, exact quantiles are calculated

mllib/src/main/scala/org/apache/spark/ml/feature/Bucketizer.scala

@@ -27,6 +27,7 @@ import org.apache.spark.ml.param._
 import org.apache.spark.ml.param.shared.{HasInputCol, HasOutputCol}
 import org.apache.spark.ml.util._
 import org.apache.spark.sql._
+import org.apache.spark.sql.expressions.UserDefinedFunction
 import org.apache.spark.sql.functions._
 import org.apache.spark.sql.types.{DoubleType, StructField, StructType}
 
@@ -46,6 +47,9 @@ final class Bucketizer @Since("1.4.0") (@Since("1.4.0") override val uid: String
    * also includes y. Splits should be of length >= 3 and strictly increasing.
    * Values at -inf, inf must be explicitly provided to cover all Double values;
    * otherwise, values outside the splits specified will be treated as errors.
+   *
+   * See also [[handleInvalid]], which can optionally create an additional bucket for NaN values.
+   *
    * @group param
    */
   @Since("1.4.0")
@@ -73,15 +77,47 @@ final class Bucketizer @Since("1.4.0") (@Since("1.4.0") override val uid: String
   @Since("1.4.0")
   def setOutputCol(value: String): this.type = set(outputCol, value)
 
+  /**
+   * Param for how to handle invalid entries. Options are skip (filter out rows with
+   * invalid values), error (throw an error), or keep (keep invalid values in a special additional
+   * bucket).
+   * Default: "error"
+   * @group param
+   */
+  @Since("2.1.0")
+  val handleInvalid: Param[String] = new Param[String](this, "handleInvalid", "how to handle" +
+    "invalid entries. Options are skip (filter out rows with invalid values), " +
+    "error (throw an error), or keep (keep invalid values in a special additional bucket).",
+    ParamValidators.inArray(Bucketizer.supportedHandleInvalid))
+
+  /** @group getParam */
+  @Since("2.1.0")
+  def getHandleInvalid: String = $(handleInvalid)
+
+  /** @group setParam */
+  @Since("2.1.0")
+  def setHandleInvalid(value: String): this.type = set(handleInvalid, value)
+  setDefault(handleInvalid, Bucketizer.ERROR_INVALID)
+
   @Since("2.0.0")
   override def transform(dataset: Dataset[_]): DataFrame = {
     transformSchema(dataset.schema)
-    val bucketizer = udf { feature: Double =>
-      Bucketizer.binarySearchForBuckets($(splits), feature)
+    val (filteredDataset, keepInvalid) = {
+      if (getHandleInvalid == Bucketizer.SKIP_INVALID) {
+        // "skip" NaN option is set, will filter out NaN values in the dataset
+        (dataset.na.drop().toDF(), false)
+      } else {
+        (dataset.toDF(), getHandleInvalid == Bucketizer.KEEP_INVALID)
+      }
+    }
+
+    val bucketizer: UserDefinedFunction = udf { (feature: Double) =>
+      Bucketizer.binarySearchForBuckets($(splits), feature, keepInvalid)
     }
-    val newCol = bucketizer(dataset($(inputCol)))
-    val newField = prepOutputField(dataset.schema)
-    dataset.withColumn($(outputCol), newCol, newField.metadata)
+
+    val newCol = bucketizer(filteredDataset($(inputCol)))
+    val newField = prepOutputField(filteredDataset.schema)
+    filteredDataset.withColumn($(outputCol), newCol, newField.metadata)
   }
 
   private def prepOutputField(schema: StructType): StructField = {
@@ -106,6 +142,12 @@ final class Bucketizer @Since("1.4.0") (@Since("1.4.0") override val uid: String
 @Since("1.6.0")
 object Bucketizer extends DefaultParamsReadable[Bucketizer] {
 
+  private[feature] val SKIP_INVALID: String = "skip"
+  private[feature] val ERROR_INVALID: String = "error"
+  private[feature] val KEEP_INVALID: String = "keep"
+  private[feature] val supportedHandleInvalid: Array[String] =
+    Array(SKIP_INVALID, ERROR_INVALID, KEEP_INVALID)
+
   /**
    * We require splits to be of length >= 3 and to be in strictly increasing order.
    * No NaN split should be accepted.
@@ -126,11 +168,26 @@ object Bucketizer extends DefaultParamsReadable[Bucketizer] {
 
   /**
    * Binary searching in several buckets to place each data point.
+   * @param splits array of split points
+   * @param feature data point
+   * @param keepInvalid NaN flag.
+   *                    Set "true" to make an extra bucket for NaN values;
+   *                    Set "false" to report an error for NaN values
+   * @return bucket for each data point
    * @throws SparkException if a feature is < splits.head or > splits.last
    */
-  private[feature] def binarySearchForBuckets(splits: Array[Double], feature: Double): Double = {
+
+  private[feature] def binarySearchForBuckets(
+      splits: Array[Double],
+      feature: Double,
+      keepInvalid: Boolean): Double = {
     if (feature.isNaN) {
-      splits.length - 1
+      if (keepInvalid) {
+        splits.length - 1
+      } else {
+        throw new SparkException("Bucketizer encountered NaN value. To handle or skip NaNs," +
+          " try setting Bucketizer.handleInvalid.")
+      }
     } else if (feature == splits.last) {
       splits.length - 2
     } else {

mllib/src/main/scala/org/apache/spark/ml/feature/QuantileDiscretizer.scala

@@ -36,6 +36,9 @@ private[feature] trait QuantileDiscretizerBase extends Params
   /**
    * Number of buckets (quantiles, or categories) into which data points are grouped. Must
    * be >= 2.
+   *
+   * See also [[handleInvalid]], which can optionally create an additional bucket for NaN values.
+   *
    * default: 2
    * @group param
    */
@@ -61,17 +64,41 @@ private[feature] trait QuantileDiscretizerBase extends Params
 
   /** @group getParam */
   def getRelativeError: Double = getOrDefault(relativeError)
+
+  /**
+   * Param for how to handle invalid entries. Options are skip (filter out rows with
+   * invalid values), error (throw an error), or keep (keep invalid values in a special additional
+   * bucket).
+   * Default: "error"
+   * @group param
+   */
+  @Since("2.1.0")
+  val handleInvalid: Param[String] = new Param[String](this, "handleInvalid", "how to handle" +
+    "invalid entries. Options are skip (filter out rows with invalid values), " +
+    "error (throw an error), or keep (keep invalid values in a special additional bucket).",
+    ParamValidators.inArray(Bucketizer.supportedHandleInvalid))
+  setDefault(handleInvalid, Bucketizer.ERROR_INVALID)
+
+  /** @group getParam */
+  @Since("2.1.0")
+  def getHandleInvalid: String = $(handleInvalid)
+
 }
 
 /**
  * `QuantileDiscretizer` takes a column with continuous features and outputs a column with binned
  * categorical features. The number of bins can be set using the `numBuckets` parameter. It is
- * possible that the number of buckets used will be less than this value, for example, if there
- * are too few distinct values of the input to create enough distinct quantiles. Note also that
- * NaN values are handled specially and placed into their own bucket. For example, if 4 buckets
- * are used, then non-NaN data will be put into buckets(0-3), but NaNs will be counted in a special
- * bucket(4).
- * The bin ranges are chosen using an approximate algorithm (see the documentation for
+ * possible that the number of buckets used will be smaller than this value, for example, if there
+ * are too few distinct values of the input to create enough distinct quantiles.
+ *
+ * NaN handling: Note also that
+ * QuantileDiscretizer will raise an error when it finds NaN values in the dataset, but the user can
+ * also choose to either keep or remove NaN values within the dataset by setting `handleInvalid`.
+ * If the user chooses to keep NaN values, they will be handled specially and placed into their own
+ * bucket, for example, if 4 buckets are used, then non-NaN data will be put into buckets[0-3],
+ * but NaNs will be counted in a special bucket[4].
+ *
+ * Algorithm: The bin ranges are chosen using an approximate algorithm (see the documentation for
  * [[org.apache.spark.sql.DataFrameStatFunctions.approxQuantile approxQuantile]]
  * for a detailed description). The precision of the approximation can be controlled with the
  * `relativeError` parameter. The lower and upper bin bounds will be `-Infinity` and `+Infinity`,
@@ -100,6 +127,10 @@ final class QuantileDiscretizer @Since("1.6.0") (@Since("1.6.0") override val ui
   @Since("1.6.0")
   def setOutputCol(value: String): this.type = set(outputCol, value)
 
+  /** @group setParam */
+  @Since("2.1.0")
+  def setHandleInvalid(value: String): this.type = set(handleInvalid, value)
+
   @Since("1.6.0")
   override def transformSchema(schema: StructType): StructType = {
     SchemaUtils.checkNumericType(schema, $(inputCol))
@@ -124,7 +155,9 @@ final class QuantileDiscretizer @Since("1.6.0") (@Since("1.6.0") override val ui
       log.warn(s"Some quantiles were identical. Bucketing to ${distinctSplits.length - 1}" +
         s" buckets as a result.")
     }
-    val bucketizer = new Bucketizer(uid).setSplits(distinctSplits.sorted)
+    val bucketizer = new Bucketizer(uid)
+      .setSplits(distinctSplits.sorted)
+      .setHandleInvalid($(handleInvalid))
     copyValues(bucketizer.setParent(this))
   }
 

mllib/src/test/scala/org/apache/spark/ml/feature/BucketizerSuite.scala

@@ -99,21 +99,32 @@ class BucketizerSuite extends SparkFunSuite with MLlibTestSparkContext with Defa
       .setOutputCol("result")
       .setSplits(splits)
 
+    bucketizer.setHandleInvalid("keep")
     bucketizer.transform(dataFrame).select("result", "expected").collect().foreach {
       case Row(x: Double, y: Double) =>
         assert(x === y,
           s"The feature value is not correct after bucketing.  Expected $y but found $x")
     }
+
+    bucketizer.setHandleInvalid("skip")
+    val skipResults: Array[Double] = bucketizer.transform(dataFrame)
+      .select("result").as[Double].collect()
+    assert(skipResults.length === 7)
+    assert(skipResults.forall(_ !== 4.0))
+
+    bucketizer.setHandleInvalid("error")
+    withClue("Bucketizer should throw error when setHandleInvalid=error and given NaN values") {
+      intercept[SparkException] {
+        bucketizer.transform(dataFrame).collect()
+      }
+    }
   }
 
   test("Bucket continuous features, with NaN splits") {
     val splits = Array(Double.NegativeInfinity, -0.5, 0.0, 0.5, Double.PositiveInfinity, Double.NaN)
-    withClue("Invalid NaN split was not caught as an invalid split!") {
+    withClue("Invalid NaN split was not caught during Bucketizer initialization") {
       intercept[IllegalArgumentException] {
-        val bucketizer: Bucketizer = new Bucketizer()
-          .setInputCol("feature")
-          .setOutputCol("result")
-          .setSplits(splits)
+        new Bucketizer().setSplits(splits)
       }
     }
   }
@@ -138,7 +149,8 @@ class BucketizerSuite extends SparkFunSuite with MLlibTestSparkContext with Defa
     val data = Array.fill(100)(Random.nextDouble())
     val splits: Array[Double] = Double.NegativeInfinity +:
       Array.fill(10)(Random.nextDouble()).sorted :+ Double.PositiveInfinity
-    val bsResult = Vectors.dense(data.map(x => Bucketizer.binarySearchForBuckets(splits, x)))
+    val bsResult = Vectors.dense(data.map(x =>
+      Bucketizer.binarySearchForBuckets(splits, x, false)))
     val lsResult = Vectors.dense(data.map(x => BucketizerSuite.linearSearchForBuckets(splits, x)))
     assert(bsResult ~== lsResult absTol 1e-5)
   }
@@ -169,7 +181,7 @@ private object BucketizerSuite extends SparkFunSuite {
   /** Check all values in splits, plus values between all splits. */
   def checkBinarySearch(splits: Array[Double]): Unit = {
     def testFeature(feature: Double, expectedBucket: Double): Unit = {
-      assert(Bucketizer.binarySearchForBuckets(splits, feature) === expectedBucket,
+      assert(Bucketizer.binarySearchForBuckets(splits, feature, false) === expectedBucket,
         s"Expected feature value $feature to be in bucket $expectedBucket with splits:" +
           s" ${splits.mkString(", ")}")
     }

mllib/src/test/scala/org/apache/spark/ml/feature/QuantileDiscretizerSuite.scala

@@ -17,10 +17,10 @@
 
 package org.apache.spark.ml.feature
 
-import org.apache.spark.SparkFunSuite
+import org.apache.spark.{SparkException, SparkFunSuite}
 import org.apache.spark.ml.util.DefaultReadWriteTest
 import org.apache.spark.mllib.util.MLlibTestSparkContext
-import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql._
 import org.apache.spark.sql.functions.udf
 
 class QuantileDiscretizerSuite
@@ -76,20 +76,33 @@ class QuantileDiscretizerSuite
     import spark.implicits._
 
     val numBuckets = 3
-    val df = sc.parallelize(Array(1.0, 1.0, 1.0, Double.NaN))
-      .map(Tuple1.apply).toDF("input")
+    val validData = Array(-0.9, -0.5, -0.3, 0.0, 0.2, 0.5, 0.9, Double.NaN, Double.NaN, Double.NaN)
+    val expectedKeep = Array(0.0, 0.0, 1.0, 1.0, 2.0, 2.0, 2.0, 3.0, 3.0, 3.0)
+    val expectedSkip = Array(0.0, 0.0, 1.0, 1.0, 2.0, 2.0, 2.0)
+
     val discretizer = new QuantileDiscretizer()
       .setInputCol("input")
       .setOutputCol("result")
       .setNumBuckets(numBuckets)
 
-    // Reserve extra one bucket for NaN
-    val expectedNumBuckets = discretizer.fit(df).getSplits.length - 1
-    val result = discretizer.fit(df).transform(df)
-    val observedNumBuckets = result.select("result").distinct.count
-    assert(observedNumBuckets == expectedNumBuckets,
-      s"Observed number of buckets are not correct." +
-        s" Expected $expectedNumBuckets but found $observedNumBuckets")
+    withClue("QuantileDiscretizer with handleInvalid=error should throw exception for NaN values") {
+      val dataFrame: DataFrame = validData.toSeq.toDF("input")
+      intercept[SparkException] {
+        discretizer.fit(dataFrame).transform(dataFrame).collect()
+      }
+    }
+
+    List(("keep", expectedKeep), ("skip", expectedSkip)).foreach{
+      case(u, v) =>
+        discretizer.setHandleInvalid(u)
+        val dataFrame: DataFrame = validData.zip(v).toSeq.toDF("input", "expected")
+        val result = discretizer.fit(dataFrame).transform(dataFrame)
+        result.select("result", "expected").collect().foreach {
+          case Row(x: Double, y: Double) =>
+            assert(x === y,
+              s"The feature value is not correct after bucketing.  Expected $y but found $x")
+        }
+    }
   }
 
   test("Test transform method on unseen data") {

python/pyspark/ml/feature.py

@@ -1155,11 +1155,6 @@ class QuantileDiscretizer(JavaEstimator, HasInputCol, HasOutputCol, JavaMLReadab
 
     `QuantileDiscretizer` takes a column with continuous features and outputs a column with binned
     categorical features. The number of bins can be set using the :py:attr:`numBuckets` parameter.
-    It is possible that the number of buckets used will be less than this value, for example, if
-    there are too few distinct values of the input to create enough distinct quantiles. Note also
-    that NaN values are handled specially and placed into their own bucket. For example, if 4
-    buckets are used, then non-NaN data will be put into buckets(0-3), but NaNs will be counted in
-    a special bucket(4).
     The bin ranges are chosen using an approximate algorithm (see the documentation for
     :py:meth:`~.DataFrameStatFunctions.approxQuantile` for a detailed description).
     The precision of the approximation can be controlled with the

sql/core/src/test/scala/org/apache/spark/sql/DataFrameStatSuite.scala

@@ -150,6 +150,10 @@ class DataFrameStatSuite extends QueryTest with SharedSQLContext {
       assert(math.abs(d1 - 2 * q1 * n) < error_double)
       assert(math.abs(d2 - 2 * q2 * n) < error_double)
     }
+    // test approxQuantile on NaN values
+    val dfNaN = Seq(Double.NaN, 1.0, Double.NaN, Double.NaN).toDF("input")
+    val resNaN = dfNaN.stat.approxQuantile("input", Array(q1, q2), epsilons.head)
+    assert(resNaN.count(_.isNaN) === 0)
   }
 
   test("crosstab") {